Text this: Continual deep reinforcement learning with task-agnostic policy distillation